The Geometric Reasoner: Manifold-Informed Latent Foresight Search for Long-Context Reasoning
Researchers have introduced The Geometric Reasoner (TGR), a novel training-free framework designed to enhance long chain-of-thought reasoning in large language models. Addressing the trade-off between computational cost and coverage quality in existing methods, TGR employs manifold-informed latent foresight search under strict memory constraints. The system scores candidate latent anchors at chunk boundaries using lightweight look-ahead estimates and soft geometric regularizers, which promote smooth trajectories and diverse exploration. By implementing chunk-wise KV cache resets, TGR maintains linear memory usage relative to chunk length. Experimental results on challenging mathematics and code benchmarks demonstrate significant improvements, with robust trajectory coverage increasing by up to 13 points on the Qwen3-8B model, measured by the area under the Pass@k curve. This performance gain is achieved with negligible computational overhead, estimated at only 1.1 to 1.3 times the baseline. The paper, submitted to arXiv under Computer Science > Machine Learning, highlights a potential breakthrough in efficient long-context reasoning without the need for extensive retraining or excessive resource consumption.
Wire timeline
The Geometric Reasoner: Manifold-Informed Latent Foresight Search for Long-Context Reasoning
Researchers have introduced The Geometric Reasoner (TGR), a novel training-free framework designed to enhance long chain-of-thought reasoning in large language models. Addressing the trade-off between computational cost and coverage quality in existing methods, TGR employs manifold-informed latent foresight search under strict memory constraints. The system scores candidate latent anchors at chunk boundaries using lightweight look-ahead estimates and soft geometric regularizers, which promote smooth trajectories and diverse exploration. By implementing chunk-wise KV cache resets, TGR maintains linear memory usage relative to chunk length. Experimental results on challenging mathematics and code benchmarks demonstrate significant improvements, with robust trajectory coverage increasing by up to 13 points on the Qwen3-8B model, measured by the area under the Pass@k curve. This performance gain is achieved with negligible computational overhead, estimated at only 1.1 to 1.3 times the baseline. The paper, submitted to arXiv under Computer Science > Machine Learning, highlights a potential breakthrough in efficient long-context reasoning without the need for extensive retraining or excessive resource consumption.
cs.AI updates on arXiv.org